run_guest_benches: clean apc_candidate_* with find, not a glob

Large guests emit thousands of per-block `apc_candidate_*` snapshots into
the shared candidates_dir (e.g. the ~7k-block guest). `rm -f dir/apc_candidate_*`
expands the glob into one argv and overflows ARG_MAX on the bench runner
("Argument list too long", exit 126), failing the whole guest-bench job
after proving has already succeeded.

Use `find -maxdepth 1 -name 'apc_candidate_*' -delete`, which deletes
in-process without building a giant argument list. `apc_candidates.json`
doesn't match the pattern, so it's still preserved.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: georgwiese

openvm-riscv/scripts/run_guest_benches.sh

@@ -56,7 +56,12 @@ run_bench() {
     # written into the shared candidates_dir on the first cache-miss run
     # and not re-created on cache hits, so this is effectively a one-time
     # cleanup per (guest, profile-input).
-    rm -f "${candidates_dir}"/apc_candidate_*
+    #
+    # Use `find -delete` rather than `rm glob`: large guests emit thousands of
+    # `apc_candidate_*` files (e.g. ~7k blocks), and the expanded glob blows
+    # past ARG_MAX ("Argument list too long"). `apc_candidates.json` doesn't
+    # match `apc_candidate_*` (no `_` after `candidate`), so it's preserved.
+    find "${candidates_dir}" -maxdepth 1 -name 'apc_candidate_*' -delete
 }
 
 # TODO: Some benchmarks are currently disabled to keep the nightly run below 6h.